Linum trains a pixel-space text-to-image model in 3.6x fewer GPU-hours than its own latent-diffusion baseline

Linum·medium signalXBlueskyLinkedInCopy link

Posted 2026-09-16 by Linum co-CEOs Sahil and Manu Chopra, JiT-DDT drops the VAE entirely and splits into an encoder predicting 64x64 structure and a decoder generating 512x512 detail, using x-prediction instead of v-prediction to handle the dimensionality problem in pixel space. Their 2.5B active-parameter model beat the 2.0B latent DiT baseline (Linum v2) using 3.6x fewer GPU-hours and 4.2x fewer samples, roughly 142M against 600M+, while producing 4x the pixels. The hardware configuration is not disclosed, so treat the GPU-hour ratio as internal and relative.